Legal work deserves fewer handoffs
A provenance-first workspace for finding what is missing, proving where it came from, and turning review decisions into follow-up work.
Reviewers lose time hunting across PDFs, notes, chats, and follow-ups.
Document, finding, source evidence, human decision, and queue stay in one workspace.
Working MVP: live dashboard, controlled agent harness, MCP access, and benchmark path.
Working MVP, live workflow surface, seeded legal review case, Claude connector available.
The messy reality of legal work
Reviewers lose time hunting across PDFs, notes, chats, and follow-ups. Documents change. Missing signatures, unsupported facts, and next-step decisions get passed between people, files, and memory.
Murdock keeps source material intact, then keeps the document, finding, source evidence, human decision, and queue in one workspace.
The point is not generic AI for law. The point is a workflow harness where the model drafts, the app validates, humans resolve, and the queue preserves follow-through.
From messy intake to reviewable work
Documents stay source-of-truth while source spans, review gates, MCP access, subagents, and benchmarked queues become the operational layer around them.
The model drafts.
The app validates.
Humans resolve.
Source material
Matter inputs stay immutable while the workspace builds an operating layer around them.
Provenance map
Claims carry page, quote, span, and document context before they become workflow state.
Review gates
Missing support, conflicts, and deadline pressure become app-owned review consequences.
Action queue
Human decisions turn recommended moves into durable case-scoped follow-through.
Agent Harness
The harness is the product argument: Murdock is not betting the case workflow on prompt obedience. It separates reading from validation, validation from consequences, and consequences from durable operations.
Model reads.
Harness validates.
Subagents execute scoped tasks.
App gates.
Queue preserves follow-through.
Draft meaning
The model receives bounded source-span bundles and drafts meaning. It does not own the matter workflow.
sourceSpanIds[] -> FindingDraftValidate claims
The compiler checks real span refs, missing support, conflicts, and materiality before anything becomes app state.
FindingDraft -> strict FindingGate risk
Review consequences are deterministic: clean facts pass, unsupported or conflicting claims become human gates.
Finding + Conflict -> ReviewGateDelegate work
Specialized subagents can take bounded real-time tasks, such as preparing a review plan or pulling evidence, without bypassing app state.
Task -> scoped agent -> resultTrack lifecycle
Resolved, ignored, queued, superseded, and done are explicit state transitions. Old issues do not vanish on refresh.
MatterOperationEvent -> current stateOperating spine
Source of truth
Original documents stay immutable; OCR and spans are traceable working copies.
Control plane
Schemas, validation, routing, review gates, queue state, and audit state live in the app.
Agent boundary
MCP exposes safe case digests, evidence, and queued tasks instead of raw database or transcript context.
Why this matters
A new upload, a changed document, or a follow-up review request does not reset the work. The system refreshes current state from the same ledger, so the UI, subagents, and connector agents see the same review memory.
The review loop, in one surface
A reviewer can see the case, the source page, the missing item, the recommended next step, and the queued follow-through without leaving the workspace.
This mockup is code-native and seeded from the real workspace vocabulary: PDF context, review gates, source links, action choices, and queued follow-through.
Immigration Case
Workspace state from I-589 intake packet, review gates, source evidence, and queued follow-through.
Signature of Attorney or Accredited Representative
The signature of the attorney or accredited representative is not provided.
14. Signature of Attorney or Accredited Representative
If you are an attorney or accredited representative, you must sign and date this form.
Signature of Attorney or Accredited Representative
Date
15. Mailing Address of Client
Provide the mailing address where you would like to receive correspondence.
Provenance
Document page, quote, and span stay attached.
Gate
Missing signature requires human review.
Queue
Accepted next step remains visible until done.
MCP
Claude reads the digest, not raw case tables.
Claude Desktop-compatible MCP
Work from Claude Desktop, with Murdock as the case source. Murdock exposes a bounded MCP connector so Claude can retrieve case review digests, source evidence, and queued action items without raw database access or transcript dumping.
Targets the mechanical 60% of the review loop
Murdock uses the harness for source finding, issue surfacing, provenance mapping, and gate setup so reviewers can spend their attention on reconciliation, next turns, and follow-through.
Legal review is rarely one clean pass. It moves in turns: inspect the packet, find the support, resolve an issue, queue the next action, then come back when the matter changes. Murdock targets the repeatable setup work in each turn. The human still owns the judgment layer: reconcile the evidence, choose the action, and carry the matter forward.
harness turn = read + find + map + gate
human turn = reconcile + decide + queue
next turn = current state + new evidence
Validation path: replay representative matters through the review loop, then report setup time reduced, judgment time preserved, correction rate, and missed-issue rate.
Built for the handoffs that slow legal work down
Less hunting. Fewer missed items. Cleaner handoffs. The goal is lower cycle time with provenance, human gates, and a clear audit trail.